We are getting half GPU-to-GPU bandwidth with Titan V cards compared to Titan X/Xp cards (measured by p2pBandwidthLatencyTest in the exact same machine and same PCIe slots), which seems to be hurting multiGPU performance. We suspect it can be a driver related problem. We are using CUDA 9.0 and
410.93 driver. Has anyone suffered the same problem?
I am also facing a similar issue. The device2device bandwidth is about 5GB/s between two Titan V. Any idea?